I am trying to use quantile regression loss function. Most of the implementations I see here, here, and here are using the error directly:
error = pred - labels
loss = max(q*error, (1-q)*error)
However, I think we should use the absolute value of the error
error = abs(pred - labels)
Practically, I have tried using non-absolute value, and the training doesn't converge, nevgative values of losses continue to decrease. I am wondering if using the absolute value in the quantile loss makes sense.
